explanation.md

Hello Andrew Trask!

I'm reading "Grokking deep learning" - I've enjoyed chapters 1-4 but Chapter 5's speed, mistakes and omitted materials have lead to a bit of a roadblock for me.

I believe I've managed to create a simple neural net, that uses gradient descent to learn but when I use the raw MNIST dataset, without transforming the pixel values from large numbers (numbers above 1) to binary (0,1) then I can't train my model, I get "double scalar" errors and all sorts of weird things (see below for examples).

I was wondering if you could look over my code, tell me where I've went wrong (tweet me: @bunsen or email: auston.bunsen@gmail.com) & if you'd like you can use this code for free, under the MIT license.

license.txt

Copyright (c) 2017 Auston Bunsen

Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all
copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
SOFTWARE.

working-version.ipynb

y-u-broken?-version.ipynb

Hey! Your training code looks good but I can't tell if the second set of testing code is correct.

Also, try reducing your alpha by an order of magnitude or two and see if results improve.

The reason I think this is likely is that if your core logic was wrong, the first network wouldn't train either. This make sme believe it's likely a tuning issue.

@iamtrask thanks! One question, why do I have to do this?

val = 1 if ord(i_f.read(1)) > 0 else 0

why doesn't the value for a pixels darkness (?) from MNIST work out of the box?

# val = ord(i_f.read(1)) why does this break it? do I need to scale this to something between 0 - 1?

it just makes the number edges sharper and easier for the network to learn